Abstract
Modern recommender systems often rely on Matrix Factorization (MF) to identify latent patterns in user-item interactions. However, these systems frequently struggle with data sparsity and the "cold start" problem. This paper proposes a novel approach that extracts "Big 5" personality traits directly from the text of user reviews to enhance recommendation accuracy. By integrating these predicted personality profiles into a Kernelized Probabilistic Matrix Factorization (KPMF) framework, the researchers achieved significant improvements in rating prediction. Experimental results demonstrate that incorporating personality-based side information can reduce Root Mean Square Error (RMSE) by up to 16% over standard MF baselines without requiring intrusive user surveys.
Introduction
Most contemporary recommendation engines focus on the properties of the items being reviewed—for instance, identifying a movie's genre or cinematography style. The authors of this study argue that this approach overlooks the most critical variable: the reviewer. Two users may award the same movie a high rating for fundamentally different reasons based on their individual dispositions. A "cinephile" might value technical mastery, while a casual viewer might prioritize entertainment value.
By capturing these inherent user traits (personality) through their unique writing styles, it is possible to build a more robust profile of user preferences. This paper introduces a method to automate this process, extracting personality signals from existing review text to guide recommendation engines.
Methodology
The proposed system utilizes a two-stage pipeline to translate raw text into actionable recommendations.
1. Personality Prediction
Using the MyPersonality dataset as a foundation, the authors trained a Ridge Regression model to predict the "Big 5" personality traits: Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism. The model relies on three categories of linguistic features:
- Linguistic Counts: Frequencies of specific parts of speech (verbs, adjectives) and punctuation.
- Sentiment and Affect: Emotional valence scores derived from the Affin lexicon.
- General Inquirer Tags: A set of 182 semantic categories used to capture the thematic "flavor" of the text.
2. Recommendation via KPMF
The core of the recommendation engine is the Kernelized Probabilistic Matrix Factorization (KPMF) framework. Unlike standard MF, KPMF allows for the integration of "side-information" through a covariance matrix. In this model, the similarity between users is calculated based on their predicted personality vectors. This similarity guides the latent factor discovery process, effectively "biasing" the model toward users with similar psychological profiles.
Datasets and Experimental Setup
The study utilized two primary datasets:
- MyPersonality Dataset: Facebook status updates from 250 users paired with formal psychological test results, used to train the personality predictor.
- IMDb Dataset: A custom-crawled dataset containing ratings and written reviews from 2,087 users across 3,500 movies, used to evaluate recommendation performance.
Key Results and Discussion
The researchers evaluated several iterations of the KPMF model against a standard Matrix Factorization baseline. The results indicate that personality traits serve as a powerful form of "guided" dimensionality reduction. While raw linguistic data is often too high-dimensional and noisy for direct use, condensing it into five psychologically-grounded scores provides a stable signal of user preference.
As shown in the results, the Standard MF baseline yielded an RMSE of 0.2262. Incorporating personality scores into the KPMF model improved this to 0.2006 (an 11.3% improvement). The most effective configuration, which combined personality scores with the underlying linguistic features, achieved an RMSE of 0.1901, representing a 16% total improvement over the baseline.
Conclusion
This research demonstrates that user personality, as manifested in writing style, is a highly predictive feature for recommender systems. By automating the extraction of these traits from reviews, the system avoids the friction of traditional psychological testing while significantly outperforming standard MF techniques. Future work may explore non-uniform weighting of personality traits—as some traits like Conscientiousness may be more predictive of taste than others—and testing the scalability of this approach on larger industrial datasets like those from Amazon or Netflix.
Summary updated